Application of deep learning to neuroimaging of psychiatric disorders with dimensional approach

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K17077
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 52030:Psychiatry-related
• Research Institution: National Center of Neurology and Psychiatry
• Principal Investigator: Yamaguchi Hiroyuki 国立研究開発法人国立精神・神経医療研究センター, 神経研究所 疾病研究第七部, 科研費研究員 (40822557)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: 深層学習 / 機械学習 / 人工知能 / 精神疾患 / 脳画像 / 次元的アプローチ

Outline of Final Research Achievements

In neuroimaging studies of psychiatric disorders, the application of deep learning techniques often involves categorical approaches to discriminate between healthy subjects and patients. However, the boundary between healthy subjects and patients remains unclear, and the clinical significance of such differentiation has yet to be firmly established. Furthermore, there is a possibility that irrelevant information is filtered out, potentially constraining the capacity of deep learning to fully manifest its capabilities.
In this study, we have adopted a dimensional approach that emphasized the correlation between symptom/behavioral indicators and their underlying biological foundation. Specifically, we have developed a deep neural network that extracts features from three-dimensional structural brain MRI images, without relying on current diagnostic labels associated with psychiatric disorders.

Free Research Field

精神医学

Academic Significance and Societal Importance of the Research Achievements

本研究により構築された深層学習モデルは、人為的バイアスを排除し、自己組織的に特徴量抽出が可能であり、これまで損失していた可能性がある脳画像に内在する情報の抽出が可能である。加えて、従来の精神疾患診断ラベルを使用しないにも関わらず、統合失調症の症状重症度や抗精神病薬の服用量と関連した特徴量を抽出でき、単に統合失調症に限定されず、多様な精神疾患の脳画像への応用の可能性を備えている。また、抽出された特徴量は生物学的指標の一つとしても期待され、将来的には精神疾患の病態解明や薬物反応性、予後予測などに活用される可能性が考えられる。